Skip to content

IsaacGym

Installation

Install IsaacGym from NVIDIA at https://developer.nvidia.com/isaac-gym. Installation instructions can be found in the package's docs folder. Python 3.8 is compatable with both IsaacGym and Sample-Factory

Install IsaacGymEnvs from https://github.com/NVIDIA-Omniverse/IsaacGymEnvs.

Running Experiments

Run IsaacGym experiments using scripts from the sf_examples.isaacgym_examples folder. Currently, we support the AllegroHand, Ant, Anymal, AnymalTerrain, BallBalance, Cartpole , Humanoid, and ShadowHand environments out of the box, and more environments can be added in train_isaacgym.py.

To run an experiment in the Ant environment: 

python -m sf_examples.isaacgym_examples.train_isaacgym --actor_worker_gpus 0 --env=Ant --train_for_env_steps=100000000  --experiment=isaacgym_ant

Multiple experiments can be run in parallel using the experiment launcher. See the experiments folder in sf_examples.isaacgym_examples for examples. To run multiple Ant and Humanoid experiments, run: 

python -m sample_factory.launcher.run --run=sf_examples.isaacgym_examples.experiments.isaacgym_basic_envs --backend=processes --max_parallel=2 --experiments_per_gpu=2 --num_gpus=1

Results

Reports

  1. We tested the IsaacGym Ant and Humanoid environments with and without recurrence. When using an RNN and recurrence, the Ant and Humanoid environments see an improvement in sample efficiency. However, there is a decrease in wall time efficiency.

  2. The AllegroHand environment was tested with and without return normalization. Return normalization is essential to this environment as it improved the performance by around 200%

Models

Environment HuggingFace Hub Models Evaluation Metrics
Ant https://huggingface.co/andrewzhang505/isaacgym_ant 11830.10 ± 875.26
Humanoid https://huggingface.co/andrewzhang505/isaacgym_humanoid 8839.07 ± 407.26
AllegroHand https://huggingface.co/andrewzhang505/isaacgym_allegrohand 3608.18 ± 1062.94

Videos

Ant Environment

Humanoid Environment

AllegroHand Environment